我是 GNN 和 PyTorch 的新手。我正在尝试使用 GNN 对系统发育数据(完全分叉、单向树)进行分类。我将 R 中的树从 phylo 格式转换为 PyTorch 数据集。以其中一棵树为例:
Data(x=[83, 1], edge_index=[2, 82], edge_attr=[82, 1], y=[1], num_nodes=83)
它有
83
节点(内部+提示,x=[83, 1]
),我为所有节点分配了0
,所以每个节点都有一个特征值0
。我构建了一个 82 X 1
矩阵,其中包含节点之间有向边的所有长度(edge_attr=[82, 1]
),我打算使用 edge_attr
表示边长度并将其用作权重。每棵树都有一个用于分类目的的标签(y=[1]
,值在 {0, 1, 2} 中)。
正如你所看到的,节点特征在我的例子中并不重要,唯一重要的是边缘特征(边缘长度)。
以下是我用于建模和训练的代码实现:
tree_dataset = TreeData(root=None, data_list=all_graphs)
class GCN(torch.nn.Module):
def __init__(self, hidden_size=32):
super(GCN, self).__init__()
self.conv1 = GCNConv(tree_dataset.num_node_features, hidden_size)
self.conv2 = GCNConv(hidden_size, hidden_size)
self.linear = Linear(hidden_size, tree_dataset.num_classes)
def forward(self, x, edge_index, edge_attr, batch):
# 1. Obtain node embeddings
x = self.conv1(x, edge_index, edge_attr)
x = x.relu()
x = self.conv2(x, edge_index, edge_attr)
# 2. Readout layer
x = global_mean_pool(x, batch) # [batch_size, hidden_channels]
# 3. Apply a final classifier
x = F.dropout(x, p=0.5, training=self.training)
x = self.linear(x)
return x
model = GCN(hidden_size=32)
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
criterion = torch.nn.CrossEntropyLoss()
train_loader = DataLoader(tree_dataset, batch_size=64, shuffle=True)
print(model)
def train():
model.train()
lost_all = 0
for data in train_loader:
optimizer.zero_grad() # Clear gradients.
out = model(data.x, data.edge_index, data.edge_attr, data.batch) # Perform a single forward pass.
loss = criterion(out, data.y) # Compute the loss.
loss.backward() # Derive gradients.
lost_all += loss.item() * data.num_graphs
optimizer.step() # Update parameters based on gradients.
return lost_all / len(train_loader.dataset)
def test(loader):
model.eval()
correct = 0
for data in loader: # Iterate in batches over the training/test dataset.
out = model(data.x, data.edge_index, data.edge_attr, data.batch)
pred = out.argmax(dim=1) # Use the class with highest probability.
correct += int((pred == data.y).sum()) # Check against ground-truth labels.
return correct / len(loader.dataset) # Derive ratio of correct predictions.
for epoch in range(1, 20):
loss = train()
train_acc = test(train_loader)
# test_acc = test(test_loader)
print(f'Epoch: {epoch:03d}, Train Acc: {train_acc:.4f}, Loss: {loss:.4f}')
看来我的代码根本不起作用:
......
Epoch: 015, Train Acc: 0.3333, Loss: 1.0988
Epoch: 016, Train Acc: 0.3333, Loss: 1.0979
Epoch: 017, Train Acc: 0.3333, Loss: 1.0938
Epoch: 018, Train Acc: 0.3333, Loss: 1.1044
Epoch: 019, Train Acc: 0.3333, Loss: 1.1012
......
Epoch: 199, Train Acc: 0.3333, Loss: 1.0965
是因为没有有意义的节点特征我们就无法使用GNN吗?还是我的实现有问题?
将所有节点特征设置为 0 是没有意义的。节点特征的意义消失了。如果节点没有特征,有一个简单的方法。这为节点创建了嵌入特征。可学习的嵌入特征用作节点的特征。
您可以随机初始化初始嵌入,然后将这些嵌入输入到 GCN 中。并且模型可以同时学习这种嵌入。
...
def __init__(self, hidden_size=32):
...
self.node_embedding = torch.nn.Embedding(
num_embeddings=self.num_nodes, embedding_dim=hidden_size)
torch.nn.init.normal_(self.node_embedding.weight, std=0.1)
...
def forward(self, x, edge_index, edge_attr, batch):
...
x = self.node_embedding.weight
x = self.conv1(x, edge_index, edge_attr)
...
...